Jev in Practice: Browser Use and Agent Workflows

The author chains Claude Code, subagents, a Playwright browser and Gmail/Drive into a local agent stack, using Jev to make file, tool and retry-or-stop micro-decisions in 70-500ms so Claude only handles what needs real reasoning.

Opus 5.5 decomposes tasks into a DAG and signs off, Haiku 5.5 runs subtasks in parallel, and Jev handles micro-branch decisions in milliseconds.

OpenRouter launches Armor in beta, using Typesafe's Jev to block out-of-policy tool calls before an agent executes them, with better performance than Claude Code's auto classifier at lower cost.

The author reconstructs the bot's state engine, six-question call, confidence gate and quote pricing, noting Jev's 70-500ms latency doesn't match microsecond HFT and that neither the paper nor the code is public.

A developer uses Jev to route messages to different bots for research and writing, with memory stored as an editable MEMORY.md file.

REAX SN92 uses small decision models instead of large ones for routing and approvals, deciding in about 26ms locally and 93ms via hosted API, and European IT firm msg is evaluating it internally.

Using Opus 5.5 as the planner and Haiku/Sonnet as executors, with Jev handling deterministic branches like tool selection and retries in milliseconds, saves a large amount of Opus tokens.

Jev now supports images, allowing text, images and structured data to be composed and queried in parallel for agent use cases like screenshot analysis and image classification.

SubQ adapted Artificial Analysis's long-context reasoning benchmark into a decision format and open-sourced 118 questions, on which Snap scores 74% and Jev 58%.

Opus writes the workflow script, Haiku scans, and Sonnet patches, while branches that need no thinking (which file, which tool, retry or stop) go to Jev for millisecond decisions.

The author shares practical tips for using Jev for computer use, with the details given only as images in the post.

A developer advises giving Jev structured context plus images, using at least 1080px resolution, escalating to stronger models when confidence is low, treating bounding boxes as one signal among several, and using a coding agent to run evals and break goals into smaller steps.